GitHub Copilot HydraFusion Max: What Is Documented

HydraFusion is a routing layer inside GitHub Copilot, not a model, and “Max” is a heavier-reasoning option added on 14 October 2026. Both are research previews, billed at each underlying model’s standard rate with no separate fee.

GitHub’s cost and quality numbers come from its own offline tests, and none of them cover Max. ClueTrove found no independent hands-on test of either option.

The short answer

What it is: it orchestrates multiple models to produce a single response; GitHub says it is “not a model”. For each prompt it picks a workflow, the models and the reasoning effort.
Status: research preview, subject to change, with no SLA and not intended for production workloads (docs).
Where: VS Code, the GitHub Copilot app and Copilot CLI.
Plans: Pro, Pro+, Business and Enterprise, subject to your organization’s model policies and preview-feature settings (14 Oct entry). The free tier is not listed.
Model pool (14 Oct update): Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna.
Billing: the standard rate of each model it uses, with no separate HydraFusion charge.
Numbers: GitHub publishes no cost or limit numbers for Max. Every performance figure traces to GitHub’s own offline tests.

What are the three workflows?

Single: one model solves the task.
Cascade: an efficient model drafts and a quality gate accepts the draft or escalates to a stronger model.
Critique: a read-only critic from a different model family reviews the drafter’s output, and the drafter revises once.

Cascade and Critique add review passes, so they take longer than a single model. That is the trade: more checking for more time and, often, more credits.

GitHub’s 14 October entry changes the defaults. Initial drafts now run at medium reasoning effort and review at high. Before, each model used its own default.

Cascade gains a judging step that asks for a repair only when the first attempt needs one. Critique now inspects the workspace with tools and runs critic models in parallel instead of leaning on text summaries.

“Max” uses the same architecture, tuned for harder tasks with more reasoning effort where depth matters. The docs say HydraFusion does not start subagents and does not undo edits from discarded drafts automatically.

What does it cost, and how do you turn it on?

Billing: per the docs, you pay each used model’s standard rate. The automatic model-selection discount does not apply. One task can use more AI credits than a single model would. The main conversation stays on one model where possible so cached tokens can still apply.
Credits: GitHub’s billing reference converts token costs to AI credits at 1 credit = $0.01. In the CLI, /usage shows credits per model.
Limits: the 14 October changelog gives no multipliers, quotas or numeric limits for either option, and ClueTrove found no billing document specific to Max.
Model choice: you cannot pick HydraFusion’s models. If none of them is available under your plan and organization policy, it does not appear in the model picker.
To enable it: in the CLI, start with –experimental or use /experimental on, then pick HydraFusion in /model. In VS Code, use version 1.140 or later with chat.copilot.hydraFusion.enabled. In the Copilot app, open Settings, then Experimental. Admins control preview features for Business and Enterprise.
Context window: the docs say it is set to a conservative value based on the smallest limit among its models, so compact a large conversation before you switch.

What do users say in GitHub’s discussion thread?

The community thread linked from the announcement (1 September to 15 October 2026) does not discuss Max. These are individual comments, not measurements.

GitHub staff: usage is priced at each underlying model’s standard rate and the models are “not subsidized”. The system is deliberately willing to spend more when that raises the chance of a correct result. No fixed roster of models is published.
Credits: one user estimated about 500 credits for planning and about 2,300 for implementation, “40% of a Pro+ package”. Another said their burn was about a tenth of what they were used to. Another said most of their spend went to one particular model. Another saw credit totals that did not match the CLI counter.
Problems mentioned: the option vanishing after organization-level model pruning, a “fusion routing returned an invalid response” error, cancelling not stopping work immediately, and difficulty seeing which tool is running.

What are GitHub’s numbers, and what are its caveats?

The 4 September launch post reports offline benchmarks against Claude Opus 5, showing the cost and quality change for HydraFusion.

Benchmark	Cost	Quality
TerminalBench 2.1	67% lower	+4.9 points
DeepSWE	36% lower	-1.5 points
CheckpointBench	65% lower	-0.1 points

Read the table this way: the big saving comes with a quality gain on only one benchmark, and quality slips slightly on the other two. Savings of 36% to 67% are large, but they apply to GitHub’s test tasks, not to your repository.

GitHub itself notes that routing was tuned across all three benchmarks and that TerminalBench 2.1 is relatively saturated. It also says the system works best for first-turn single-prompt tasks, and that names, availability and behavior may change.

The comparison is with Opus 5, not Opus 5.5, and none of these figures cover Max.

Is there independent evidence?

ClueTrove found no hands-on test or benchmark replication, for HydraFusion or for Max. What exists is commentary on GitHub’s numbers:

VentureBeat (4 September) notes that quality matched or exceeded Opus 5 in one of the three benchmarks. It quotes a developer objecting that you pay the cheap model on every cascade request.
Pondero (8 September) models a hypothetical bill, which is a calculation, not a measurement.
InfoQ (13 September) relays GitHub’s figures.
A DEV Community post (12 October) logged a routing decision at 251 ms on CLI 1.0.89, with no quality results, and predates Max.

Where do the sources disagree?

The launch post says access is included on all Copilot plans. The 14 October entry names the four paid plans. GitHub’s docs and the 14 October entry do not mention a free plan, so treat free-tier access as unconfirmed.
Model names differ between GitHub posts: GPT-5.6 Sol and Luna in the September post, GPT-6 Astra, Sol and Luna on 14 October. This article uses the latest.

Who should try it, and who should wait?

Try it if you already pay for a listed plan and have a bounded task where a second model’s review would help. Skip it for production workloads or for budgets you cannot let drift.

Try it on one bounded task and check credits per model with /usage.
Keep it out of production workloads while it is a preview without an SLA.
Inspect the real output; see when passing tests hide an unfinished task.
Judge it by cost per accepted task, not token price: why a cheaper model can cost more.

Questions and answers

What is HydraFusion Max?

An option inside HydraFusion tuned for harder tasks with higher reasoning effort, according to GitHub. It is a research preview.

Is HydraFusion a new model?

No. The docs say it is not a model. It picks a workflow, models and effort per prompt.

How is it billed?

At each used model’s standard rate with no separate fee, so a task can cost more than one model would. GitHub publishes no Max-specific figures.

Can I choose which models it uses?

No. It picks them per prompt from its pool, limited to what your plan and organization policy allow.

Should I use it in production?

No. The docs say it has no SLA and is not intended for production workloads.

Terms in plain language

Routing layer: software that decides which model handles each request.
Reasoning effort: how much thinking a model does before it answers.
SLA: a written promise about service availability.
AI credits: the unit GitHub uses to bill model usage.

Sources

Facts last checked against the linked pages on 15 October 2026 and re-checked against the GitHub changelog and docs on 29 October 2026. Preview features can change.

Part of the guide: AI Model Pricing and API Cost Guides